This page describes runtime egress for each supported search engine, split into standard search and grounding (agent search). Each kind has different request shapes and data-leave characteristics.

Credentials and env vars: Search Engine Setup (Infra). Visibility: Activation Reference. Isolation boundary: Search Proxy.

Shared path when Search Proxy is enabled

When Search Proxy is disabled, engines that still support legacy mode call providers directly from assistants-core. Brave and Perplexity have no legacy path.

Standard search

Standard search sends a query (plus optional admin/AI knobs) to a search API and receives a list of results. Parameter exposure (fixed vs AI-overridable knobs) is a Spaces concern documented under Technical — Search Engine Configuration; it only applies here, not to grounding.

Google Search

• Egress destination: Google Custom Search JSON API.

• With Search Proxy: assistants-core → Search Proxy → Google.

• Request contains query, CSE id / API key (held on Search Proxy when enabled), optional locale/date/site filters.

• Response: URLs, titles, snippets — typically requires a Web Page Reader afterward.

• Google Custom Search JSON API has announced lifecycle limits; plan migration where needed.

Brave Search (proxy-only)

• New standard provider. Requires Search Proxy.

• Egress: Brave Search API from the Search Proxy service.

• Rich snippets; full page crawl often optional.

• No legacy direct path from assistants-core.

Perplexity (proxy-only)

• New standard provider. Requires Search Proxy.

• Egress: Perplexity Search API from Search Proxy.

• Configurable content extraction depth / token limits and filters.

• No legacy direct path from assistants-core.

Customized API Search

• Not routed through Search Proxy.

• assistants-core calls the configured customer/vendor REST endpoint directly (subject to cluster egress policy for that destination).

• Response must match the Web Search results schema.

Grounding (agent search)

Grounding does not use standard search parameter exposure. An upstream agent or grounded model performs research; Search Proxy (when enabled) exposes an agent-search capability that returns an opaque answer for assistants-core to parse into structured results.

Grounding with Bing

• Destination: Azure AI Foundry project / Bing grounding agent.

• With Search Proxy: assistants-core → Search Proxy agent-search → Azure / Bing.

• Modes: auto-provisioned agent vs pre-configured agent id.

• Default: content already in the answer → Web Page Reader usually skipped (requires_scraping configurable).

• Higher typical latency than standard search APIs.

Grounding with VertexAI

• Destination: Vertex AI Gemini with Google Search grounding.

• With Search Proxy: assistants-core → Search Proxy agent-search → Vertex / Google.

• May include grounding redirect URLs; optional redirect resolution to canonical sites.

• Default: content in the answer → Web Page Reader usually skipped.

Data-leave summary

Query strings can include conversation context — see the Web Search root warning and Argument Screening.